SPECT hybrid approach combining conventional and deep features for early detection of Parkinson's disease.

The early diagnosis of Parkinson's disease (PD) using SPECT imaging continues to be challenging due to the subtle dopaminergic deficits present in the early stages of the disease. This study proposes a novel hybrid approach combining conventional and deep learning features to improve PD classification, and applies it to reclassify scans without evidence of dopaminergic deficit (SWEDD) cases. We used SPECT images from early PD patients and healthy controls (HC) and extracted SBR metrics, morphometric, and deep learning features. Our multi-stage feature selection pipeline employed near-zero variance filtering, ANOVA F-test analysis, correlation-based feature elimination, and Random Forest importance scoring. We evaluated multiple machine learning algorithms and selected Linear Discriminant Analysis as the optimal classifier, then applied this model to reclassify 79 SWEDD cases. Feature selection reduced 79 significant features to 15 optimal features: 1 SBR metric (6.7%), 7 morphometric (46.7%), and 7 deep features (46.7%). The hybrid Linear Discriminant model achieved the best performance, outperforming individual feature approaches with 97.40% test accuracy, 96.25% sensitivity, 98.65% specificity, and 99.59% AUC. Statistical analysis revealed morphometric features had the highest mean importance (0.0699 ± 0.0539), followed by deep (0.0400 ± 0.0570) and SBR features (0.0206 ± 0.019). SWEDD reclassification identified 5 cases (6.3%) with imaging patterns consistent with early PD, while 74 cases (93.7%) maintained HC characteristics. This study presents a proof-of-concept demonstration of the effectiveness of integrating conventional measures with deep learning techniques for improving the early diagnosis of PD, while offering new insights into SWEDD case reclassification.